Giving voice to digital twins: How LLMs build human knowledge graphs

The convergence of the Internet of Things, artificial intelligence, and semantic technologies is reshaping how digital systems model and reason about human behavior. Human Digital Twins (HDTs) offer dynamic user representations by integrating multimodal data from physiological, behavioral, and linguistic sources. However, most existing HDTs perform low-level data fusion, lacking the semantic consistency needed for coherent reasoning and personalization. This paper proposes a novel pipeline for constructing Personal Knowledge Graphs (PKGs), i.e. adaptive, formal representations of user knowledge, that combine the generative capabilities of Large Language Models (LLMs) with ontology-based validation. The pipeline includes triplet extraction from language data, semantic alignment through embeddings, and ontology-driven graph construction. A comparative analysis between ontology-free and ontology-guided PKG generation demonstrates that semantic grounding substantially reduces spurious relationships and improves reasoning accuracy, establishing ontological validation as essential for reliable knowledge extraction. By bridging symbolic and generative paradigms, the proposed approach advances knowledge-driven HDTs that are proactive, interpretable, and resilient.